onehotdecode

Decode probability vectors into class labels

Syntax

A = onehotdecode(B,classes,featureDim)

A = onehotdecode(B,classes,featureDim,typename)

Description

A = onehotdecode(B,classes,featureDim) decodes each probability vector in B to the most probable class label from the labels specified by classes. featureDim specifies the dimension along which the probability vectors are defined. The function decodes the probability vectors into class labels by matching the position of the highest value in the vector with the class label in the corresponding position in classes. Each probability vector in A is replaced with the value of classes that corresponds to the highest value in the probability vector.

A = onehotdecode(B,classes,featureDim,typename) decodes each probability vector in B to the most probable class label and returns the result with data type typename. Use this syntax to obtain decoded class labels with a specific data type.

Examples

Encode and Decode Labels

Use the onehotencode and onehotdecode functions to encode a set of labels into probability vectors and decode them back into labels.

Create a vector of categorical labels.

colorsOriginal = ["red" "blue" "red" "green" "yellow" "blue"];
colorsOriginal = categorical(colorsOriginal)

colorsOriginal = 1x6 categorical
     red      blue      red      green      yellow      blue 

Determine the classes in the categorical vector.

classes = categories(colorsOriginal);

One-hot encode the labels into probability vectors by using the onehotencode function. Encode the probability vectors into the first dimension.

colorsEncoded = onehotencode(colorsOriginal,1)

colorsEncoded = 4×6

     0     1     0     0     0     1
     0     0     0     1     0     0
     1     0     1     0     0     0
     0     0     0     0     1     0

Use onehotdecode to decode the probability vectors.

colorsDecoded = onehotdecode(colorsEncoded,classes,1)

colorsDecoded = 1x6 categorical
     red      blue      red      green      yellow      blue 

The decoded labels match the original labels.

Decode Probability Vectors into Most Probable Classes

Use onehotdecode to decode a set of probability vectors into the most probable class for each observation.

Create a set of 10 random probability vectors. The vectors express the probability that an observation belongs to one of five classes.

numObs = 10;
numClasses = 5;

prob = rand(numObs,numClasses);

tot = sum(prob,2);
prob = prob./tot;

Define the set of five classes.

classes = ["Red" "Yellow" "Green" "Blue" "Purple"];

Decode the probabilities into the most probable classes. The probability vectors are encoded into the second dimension, so specify the dimension containing encoded probabilities as 2. Obtain the most probable classes as a vector of strings.

result = onehotdecode(prob,classes,2,"string")

result = 10x1 string
    "Red"
    "Yellow"
    "Yellow"
    "Green"
    "Yellow"
    "Blue"
    "Green"
    "Yellow"
    "Red"
    "Red"

Input Arguments

B — Probability vectors

numeric array

Probability vectors to decode, specified as a numeric array.

Values in B must be between 0 and 1. If a probability vector in B contains NaN values, then the function decodes that observation to the class that has the largest probability that is not NaN. If an observation contains only NaN values, the function decodes that observation to the first class label in classes.

Data Types: single | double

classes — Classes

cell array | string vector | numeric vector | character array

Classes, specified as a cell array of character vectors, a string vector, a numeric vector, or a two-dimensional character array.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64 | string | cell

featureDim — Dimension containing probability vectors

positive integer

Dimension containing probability vectors, specified as a positive integer.

Use featureDim to specify the dimension in B that contains the probability vectors. The function replaces each vector in B along the specified dimension by the element of classes in the same position as the highest value along the vector.

The dimension of B specified by featureDim must have length equal to the number of classes specified by classes.

typename — Data type of decoded labels

'categorical' (default) | character vector | string scalar

Data type of decoded labels, specified as a character vector or a string scalar.

Valid values of typename are 'categorical', 'string' and numeric types such as 'single' and 'int64'. If you specify a numeric type, classes must be a numeric vector.

Example: 'double'

Data Types: char | string

Output Arguments

A — Decoded class labels

categorical array (default) | string array | numeric array

Decoded class labels, returned as a categorical array, a string array, or a numeric array.